henhaobio
Sign in

Scientific ReportsBiochemistryPeer-reviewed journal article

Medium-sized protein language models perform well at transfer learning on realistic datasets

Luiz Carlos VieiraiD, Morgan L. Handojo, Claus O. WilkeiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Protein language models (pLMs) can offer deep insights into evolutionary and structural properties of proteins. While larger models, such as the 15 billion parameter model ESM-2, promise to capture more complex patterns in sequence space, they also present practical challenges due to their high dimensionality and high computational cost. We systematically evaluated the performance of various ESM-style models across multiple biological datasets to assess the impact of model size on transfer learning via feature extraction. Surprisingly, we found that larger models do not necessarily outperform smaller ones, in particular when data is limited. Medium-sized models, such as ESM-2 650M and ESM C 600M, demonstrated consistently good performance, falling only slightly behind their larger counterparts--ESM-2 15B and ESM C 6B--despite being many times smaller. Additionally, we compared various methods of compressing embeddings prior to transfer learning, and we found that mean embeddings consistently outperformed other compression methods. In summary, ESM C 600M with mean embeddings offers an optimal balance between performance and efficiency, making it a practical and scalable choice for transfer learning in realistic biological applications.

In-depth summary

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 3 (8 May 2025) · CC BY 4.0

Medium-sized protein language models (e.g. ESM C 600M) with mean embeddings nearly match ESM-2 15B and ESM C 6B in transfer learning on mutational scanning and PISCES data.

Protein language models (pLMs) have grown to 15 billion parameters (ESM-2) and 98 billion (ESM3), but large models produce high-dimensional embeddings and are expensive to run, and the paper cites work doubting that scale is the best route. The work benchmarks whether parameter count and embedding compression change accuracy when pLM embeddings serve as features for supervised prediction on realistic, often modest-sized, datasets.

Key results

  1. ShownMean pooling of per-residue embeddings was the best general compression method. On average it beat max pooling, the BOS token, PCA, kernel PCA and iDCT, although some alternatives did slightly better on individual DMS datasets.

    Using last-layer ESM-2 150M embeddings with LassoCV, mixed-effects models (compression as fixed effect, dataset as random effect) gave mean pooling a gain in hold-out variance explained of 5 to 20 percentage points over alternatives on DMS data and 20 to 80 points on 12 PISCES targets. Max pooling, iDCT or PCA were slightly better on some DMS sets. · Mean embeddings outperform all other compression…; Per-token embeddings compression; Statistical Analysis

  2. ShownGains from scaling were small. Medium-sized models came close to, or beat, much larger ones, and the newer ESM C family outperformed the largest ESM-2 model.

    On 36 DMS datasets (proteins up to 1,022 residues), ESM-2 150M slightly beat ESM-2 3B despite being twenty times smaller. Every ESM C model (300M, 600M, 6B) beat ESM-2 15B, and ESM C 6B had only a minor edge over ESM C 600M. On 12 PISCES targets, all models with 150M or more parameters performed comparably. · Moderately sized models perform well in transfer…; Abstract

  3. SuggestedLoRA fine-tuning of ESM-2 models showed the same size pattern, levelling off near 650M parameters. Fine-tuning gave systematically higher accuracy than feature extraction, and the two sets of results were correlated.

    155 fine-tuned models: five ESM-2 sizes (8M to 3B) on 31 DMS datasets with under 15,000 samples and proteins under 800 residues, one run each. ESM-2 3B, 4.6 times larger, did not beat ESM-2 650M and was sometimes slightly worse (supplementary figures). · Moderately sized models perform well in transfer…; Fine-tuning ESM-2 models; Discussion; Limitations

  4. SuggestedTraining-set size limits what large embeddings can deliver. All models lost accuracy on small subsets, and ESM-2 15B overtook smaller models only once LassoCV retained more of its features.

    Three DMS datasets (single, double and multiple mutants) were downsampled from 100 observations to full size. With enough data, all models except ESM-2 8M/35M and AMPLIFY 120M/350M performed comparably. ESM-2 650M and ESM C 300M/600M used about half as many LassoCV features as ESM-2 15B, with comparable performance. · Sample size drives transfer learning performance…

  5. SuggestedAcross DMS datasets, protein length explained more of the variation in predictability than dataset size. Long proteins, and viral ones in particular, were predicted worst. The last embedding layer was usually the best choice.

    In single-predictor regressions on 40 DMS datasets (ESM C 600M), sample size explained 0.004 of the variation and protein length 0.083. Proteins over 1,022 residues were linked to lower scores, and viral proteins had some of the lowest. In a layer-wise LassoCV scan of all 30 layers of a 300M model, the last layer was most often best and otherwise the penultimate. · Transfer learning is limited by data quality; Layer-wise analysis in protein language model; Medium-size models perform well in transfer…

Methods, data and reproducibility

Last-hidden-layer embeddings from ESM-2 (8M to 15B), ESM-1v, ESM C and AMPLIFY models were compressed (mean or max pooling, BOS token, PCA, kernel PCA, iDCT) and used as features in LassoCV regression with 5-fold cross-validation, to predict DMS fitness values and PISCES-derived protein properties. Five ESM-2 models were also LoRA fine-tuned with a regression head.

  • Deep mutational scanning datasets from a prior curation (stability, enzyme activity, binding, MIC, viral replication, growth, etc.); 36 used for the model-size comparison, 31 for fine-tuning · multiple (human, bacterial, yeast and viral proteins) · n = 41 datasets (40 or 36 in some analyses)
  • PISCES proteins with known structures (at most 50% pairwise similarity), 64 to 1022 residues; 12 targets: 6 physicochemical properties, 3 amino acid frequencies, 3 secondary-structure frequencies · n = 23,487 sequences
Methods & tools
ESM-2 (8M to 15B), ESM-1v, ESM C (300M, 600M, 6B), AMPLIFY (120M, 350M), scikit-learn (LassoCV, PCA, KernelPCA), lme4, multcomp, Biopython DSSP, Peptides 0.3.2, LoRA via Parameter-Efficient Fine-Tuning, PyTorch, iDCT compression from PROST
Software & environment
ESM C run in float32 with autocasting off; AMPLIFY embeddings LayerNorm-normalized, max_length raised to 5000, sequences processed singly without padding; Peptides 0.3.2; LoRA rank 4, alpha 32. Authors advise checking GPU embeddings against CPU float32 runs.
Assay conditions & compute
Fine-tuning ran on AMD Instinct MI100 GPUs. ESM-2 15B could not be fine-tuned on an NVIDIA H100 even with LoRA. The authors note their large (billion-parameter) class tends to need multiple high-end GPUs. No runtimes or memory figures are given.
Reproducibility gaps
ESM C 6B required beta-program enrollment and an API token (not public at writing). Default ESM C precision settings and batched AMPLIFY inference gave non-reproducible embeddings, fixed only by the edits described. DMS data are reused from a prior curation listed in a metadata CSV.
Limitations (6)
  • Only ESM-2 and related models (ESM-1v, ESM C, AMPLIFY) were tested. ProtTrans, ProtFlash and generative models such as ProGen were left out, though the authors expect including them would not change the conclusions. · stated by the authors
  • Only simple compression schemes (mean, max, BOS token, PCA, kernel PCA, iDCT) were compared; compression methods trained on the data could outperform mean pooling. · stated by the authors
  • Fine-tuning was not explored systematically because of compute. ESM-2 15B could not be fine-tuned even on NVIDIA H100 GPUs with LoRA, and ESM-2 3B could not be fine-tuned on the largest datasets or longest proteins. · stated by the authors
  • ESM C 6B was reachable only through a beta-access API with a 2,048-residue limit, so it has no results for the HIS dataset (500,000 sequences) or POLG Sun2014 (3,033 residues). The long-protein analysis rested on four proteins and was inconclusive. · stated by the authors
  • The compression comparison used only ESM-2 150M embeddings. Whether mean pooling is also best for the other model families, including ESM C 600M in the final recommendation, was assumed rather than tested. · noted in this summary
  • Each of the 155 fine-tuned models was trained once on a single 80/20 split, with the epoch picked from training versus validation R curves. Small size differences, such as 3B vs 650M, therefore come with no replicate variance. · noted in this summary

Read this if you use protein language model embeddings as features for supervised property or variant-effect prediction and want guidance on model size, pooling, and settings that keep embeddings reproducible across hardware.

Citations
42
FWCI
11.60
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISSerrano '23Cheng '24Senoner '25Çelik '25Li '24Naderializa '24
74% semantic matchBioinformatics
Yaiza Serrano, Sergi Rodà, Vı́ctor Guallar+1 more·bioRxiv·Aug 4, 2023· 8 citations